5 posts tagged with "OAuth"

Agents behave non-deterministically, they now hold increasingly powerful permissions, and the blast radius keeps growing when they fall under external Agent Goal Hijacking or simply misfire during ordinary work.

That is why kill switches keep coming up as one way to stop destructive changes made by an agent. The surrounding ecosystem is filling in too, with Anthropic shipping Inference hooks as a new capability.

At the start of the year, in AI Security Challenges in 2026, I laid out the challenges the industry as a whole would likely face over the course of the year. Half a year has now passed, and many of the topics I described back then as needing to "wait for future technological maturation" have already reached a stage where we can start taking action. Okta Cross-App Access has become a service ready for production deployment, and Anthropic has adopted it. Sandboxes for agents are becoming standardized, and frontier models are beginning to surpass humans in the application security domain. As we enter the second half of 2026, this article summarizes the latest state of the AI Security industry and lays out where those responsible for AI Security at a typical organization—one that is presumably driving agent adoption—should start over the next six months.

The year 2025 witnessed a continuous cycle of emerging and evolving AI/LLM technologies. Across the industry, various security measures for AI have been advancing. In 2026, AI adoption is expected to expand further, bringing new technologies and demanding corresponding security measures. This article reviews the major topics from 2025, breaks down the anticipated industry-wide challenges in AI Security for 2026 into concrete action items, and summarizes security measures for safely using and providing AI.

MCP, an extension layer for AI and LLM systems, now underpins many AI agents. At the same time, multiple attack vectors have been identified that build on this new AI and LLM context. The best known is MCP tool poisoning, where an agent is tricked into running harmful actions. This was discovered because of the combination of nondeterministic LLM behavior and the new concept of shared context.

While new attack methods emerge, long standing attack vectors are also being exposed by unsafe parts of the MCP specification.

As of July 2025, the Model Context Protocol (MCP) has evolved through three major versions (2024-11-05, 2025-03-26, and 2025-06-18). Today, MCP has become an integral part of our daily development workflow, with developers launching MCP servers via npx commands, uv commands, or Docker containers from MCP clients like Cursor, Claude Code, and VS Code, while also connecting to remote MCP servers.